Space Agent is a browser-first AI agent platform where the agent doesn’t just answer questions — it actively reshapes the interface around you. Ask for a new page, a widget, or an entire workflow and the agent can build it directly into the running workspace while you work, extending your Space toward whatever you can imagine. Created by Agent Zero, Space Agent runs in the browser layer itself, giving the agent direct access to the same framework, modules, and UI it is reshaping.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/agent0ai/space-agent/llms.txt
Use this file to discover all available pages before exploring further.
Why Space Agent Is Different
Unlike traditional AI assistants confined to a fixed chat interface, Space Agent lives in the frontend runtime. Whether you open it in a browser tab or through the desktop app, the agent can work directly with the same framework and UI it modifies — no round-trips to a separate backend required for most operations.Agent Reshapes the UI
The agent builds pages, tools, and widgets straight into the running workspace as you work — the interface is never fixed.
Browser-First Runtime
Application logic lives in the browser layer. The Node.js server is thin infrastructure for auth, APIs, module delivery, and fetch proxying.
SKILL.md Capabilities
New agent capabilities can live in plain-text
SKILL.md files the agent can write and extend itself — no redeployment needed.Token-Efficient Execution
No bulky tool-call JSON. The agent stays in plain text and plain JavaScript inside the same message when action is needed.
The Agent Lives in the Frontend
Space Agent’s primary runtime is the browser app inapp/. The server in server/ is deliberately thin — it handles page shells, authentication, APIs, module delivery, app-file access, and outbound fetch proxying. New behavior belongs in the frontend unless a backend boundary is required for security, shared-data integrity, cross-user isolation, or runtime stability.
This means the agent can act through browser state and narrow backend contracts instead of moving application logic onto the server. The result is a system that is faster to extend, easier to reason about, and naturally suited to AI-driven development.
Browser-First Architecture
The runtime surfaces you interact with map directly to how the layers are split:| Surface | Purpose |
|---|---|
/ | Authenticated app shell — the main agent workspace |
/admin | Firmware-only admin shell, clamped to L0 (maxLayer=0) |
/login | Public sign-in shell |
/enter | Recovery-safe launcher shell |
/api/<endpoint> | Backend endpoint surface |
/api/proxy | Authenticated outbound fetch proxy |
/mod/<author>/<repo>/... | Browser module delivery |
Layered Customware: L0, L1, L2
Space Agent uses three logical layers for its module resolver, giving you a clean separation between core firmware and user-owned work:- L0 — Firmware: Immutable during normal runtime. The core system lives here and is never overwritten by user or group customization.
- L1 — Group customware: Shared tools, workflows, and behavior your team can develop and publish to one another.
- L2 — User customware: Personal work that lives entirely in your own layer without affecting anyone else.
CUSTOMWARE_PATH is configured, writable L1 and L2 roots live at CUSTOMWARE_PATH/L1 and CUSTOMWARE_PATH/L2, keeping mutable state outside the source checkout so production updates never clobber your work.
Text-Based Skills
New capabilities can live in simpleSKILL.md files. The agent can write and extend these files itself in plain text, meaning the system can grow its own abilities without requiring code changes, recompilation, or redeployment. Skills become part of the agent’s knowledge the moment they are written.
Git-Backed Time Travel
Every writable L1 (group) and L2 (user) root can maintain its own local Git history repository, debounced and auto-committed as changes accumulate. When something breaks, you can roll back a user’s or group’s customware without affecting anyone else. Admin mode provides a stable control plane that is always accessible even when user-layer changes cause problems.Token-Efficient Execution
Most agent frameworks communicate intent through verbose JSON tool-call payloads. Space Agent lets the agent stay in plain text and plain JavaScript inside the same message when action is needed — no wrapping, no schema overhead, less token consumption per operation.Three Deployment Modes
Desktop App
One bundled app, no terminal required. Download from GitHub Releases and run on macOS, Windows, or Linux.
Personal Server
Run
node space serve for a local or single-user setup with optional no-login mode.Team / Production
node space supervise delivers zero-downtime auto-updates for shared team deployments.Personal to Hierarchical
Space Agent scales from a single personal assistant all the way to a hierarchical system of users and groups. Users build in their own L2 layer without affecting teammates. Groups share tools, workflows, and behavior through L1 when they are ready. The_admin group grants access to the admin shell and system management commands.
Puzzle-Piece Modularity
The core stays small. Most of Space Agent is built from modular pieces — Spaces, Widgets, Skills, and modules — that can be added, removed, or swapped cleanly instead of being welded into one rigid application. This is what makes it possible for the agent to extend the system from within.Explore Further
Quickstart
Clone the repo and have a working Space Agent server running in under five minutes.
Deployment
Desktop app, personal server, and production team deployment — all three modes explained.
CLI Reference
Full command surface with built-in help for each command, runtime parameter schema, types, and defaults.
Architecture Overview
Explore the browser-first runtime model, server surfaces, and how the layers connect.